active insulin infusion using optimal and derivative ... · embedded computation and emerging...
TRANSCRIPT
Active Insulin Infusion Using Optimal and Derivative-Weighted Control
Z-H Lam1, K-S Hwang1, J-Y Lee1, J. Geoffrey Chase2, and Graeme C. Wake3
University of Canterbury
Dept of Mechanical Engineering
Private Bag 4800
Christchurch, New Zealand
Email: [email protected]
1 Research Assistant, 2 Sr. Lecturer Mechanical Eng., 3 Professor, Applied Mathematics
1
ABSTRACT
Close control of blood glucose levels significantly reduces vascular complications in Type I
diabetes. A control method for the automation of insulin infusion that utilizes emerging
technologies in blood glucose biosensors is presented. The controller developed provides
tighter, more optimal control of blood glucose levels, while accounting for variation in patient
response, insulin employed and sensor bandwidth. Particular emphasis is placed on controller
simplicity and robustness necessary for medical devices and implants.
A PD controller with heavy emphasis on the derivative term is found to outperform the
typically used proportional-weighted controllers in glucose tolerance and multi-meal tests.
Simulation results show reductions of over 50% in the magnitude and duration of blood
glucose excursions from basal levels. A closed-form steady state optimal solution is also
developed as a benchmark, and results in a flat glucose response. The impact and tradeoffs
associated with sensor bandwidth, sensor lag and proportional versus derivative based control
methods are illustrated. Overall, emerging blood glucose sensor technologies that enable
frequent measurement are shown to enable more effective, automated control of blood
glucose levels within a tight, acceptable range for Type I and Type II diabetic individuals.
Keywords: Diabetes, Optimal Control, Automated, Insulin Infusion, Sensor Lag
1. INTRODUCTION
If not managed properly, diabetes can lead to complications such as nerve damage, brain
damage, amputation and eventually death. Diabetes related complications are a worldwide
2
epidemic with high medical, economic and social costs [1-3]. Tight control of blood glucose
levels has also been shown to reduce the mortality of diabetic, and non-diabetic, intensive
care unit patients by up to 50% [4]. Diabetic individuals monitor food intake and daily
activity to maintain blood sugar levels at an adequate level. For ease of management, subjects
are encouraged to stick to strict routines and diets to minimize manual monitoring and
injections, reducing intervention and difficulty. This regime can lead to severe limitation of
the subjects’ lifestyle, an “institutional” psychology, and the difficulty of consistently
maintaining a strict daily regimen over several years.
Though devices that can measure glucose level and administer insulin exist, they do just that –
measure and inject - with no automated interface between the two. A diabetic individual is
required to carry out the procedures manually, introducing drudgery and the error due to
human miscalculation and limitations. Hence, a system that automatically monitors and
controls a diabetic individual’s blood sugar level allows the patient to more fully engage in
the “normal” routines of life with reduced risk of long-term adverse end-results.
Current treatments for Type I diabetes involve monitoring the plasma glucose level and
injecting insulin as needed. Normally, patients follow a strict regimen to prevent
complications, however the effectiveness of this regimen is a function of the patient's intuition
and experience. A typical day for a diabetic individual might involve injecting long-acting
insulin approximately three times and injecting rapid-acting insulin before meals, to reduce
the post-prandial blood glucose spike. The patient is required to make intuitive decisions
when they deviate from normal diet or exercise patterns, and modify their regimen to suit the
irregularity. As a result, error is introduced and control is often not optimal, even for
conscientious patients.
3
Most commonly available glucose sensing devices operate by measuring the blood glucose
content of a small finger-prick blood sample, an irritating method upon frequent use. As a
result, some diabetic individuals measure blood sugar as infrequently as once per day, or less.
However, recent advances have resulted in semi-invasive systems such as the GlucoWatch
Biographer from Cygnus [5]. This device offers sampling rates up to one reading every 20
minutes, and can measure and store data continuously for up to 12 hours before new sensor
pads are required.
This research develops optimal and robust control algorithms for the automated infusion of
insulin to Type I and Type II diabetic individuals. Wireless technology, insulin pumps,
embedded computation and emerging non-invasive glucose monitoring systems may be
interconnected and a closed loop control system realized. Ultimately, an effective control
system should be able to automate over 95% of a diabetic individual’s regular insulin care,
freeing them from significant stress and to better handle any exceptions that arise. The
potential for future implementation(s) of an implanted artificial pancreas is investigated in
terms of the feasibility of the interface and the tradeoffs between control performance, and
sensor and actuator limits.
Current diabetes management typically seeks to limit blood glucose levels to be less than 5.8-
6.0 mmol/L, however tighter control to the basal level of 4.5 mmol/L would significantly
limit the damage that can result from long-term exposure to elevated blood glucose levels.
Such relaxed higher blood glucose levels are considered acceptable, as blood glucose
management without automation does not typically deliver the data, or the ability to
constantly modify insulin infusion rates, necessary for tighter control. In addition, regular,
4
automated blood glucose measurement provides the consistent volume of data necessary for
such tight control, in contrast to the sometimes infrequent and inconsistent self-measurement
reported by clinicians. Automation can have a significant impact on the treatment of juveniles
with Type I diabetes where children have less ability to fully manage the disease and, as a
result, the stress and difficulty of management can fall on parents and others. In all cases, an
automated approach can provide clinical staff with a far greater abundance, and consistency,
of data, and the resulting ability to provide better treatment and feedback.
Whether it is in the area of understanding, modeling or managing diabetes [6,7] years of
research in this area has led to no shortage of theoretical solutions [8-17]. However, due to
either the complexity of the proposed implementation, current technological limitations,
models that are not physiologically verified, lack of required data, or the cost/complexity of
realizing the results, these solutions are not yet practicable.
Several researchers have examined the analysis and automation of insulin administration as
reviewed by Lehman et al. [18]. Many of the systems presented use control as a means of
providing clinical advice or testing the efficacy of a new protocol [19-24]. A more complex,
higher performance real-time control example uses model predictive control on a 19th order
model of the glucose-regulatory system resulting in a 40% peak reduction and 23% reduction
in settling time to basal level [25]. Optimal control using grid search theory, robust H-infinity
control, and variable structure controllers have also been studied, each using different models
[11, 24-28]. In each case, the focus has been on controlling absolute blood glucose excursion
rather than slopes. The models used typically require either patient specific parameters that
are not generically available, and/or knowledge of glucose or exercise inputs that would not
5
be known a priori. Finally, none study the impact of more frequent measurement enabled by
recent advances in sensing technology or the potential for improved results.
The following sections present the physiological model employed, the optimal control
solution and the heavy derivative PD controllers developed. Subsequent sections examine the
impact of sensor bandwidth, sensor lag and other forms of error.
2. MODEL OF THE GLUCOSE REGULATORY SYSTEM
Models of the glucose regulatory system can be categorised as either comprehensive or
simple models. Comprehensive models, though they are very accurate in regimen evaluation,
are generally unsuited for real-time control, requiring several time points of input to generate
the insulin infusion profile. Additionally, they are not generic requiring patient-specific data
and known glucose inputs. The aim of this research is to develop control schemes based on
models that capture the essential system dynamics, do not require unavailable data, and are
applicable to a wider variety of subjects. Simple models capture these essential dynamic
behaviours, providing a more suitable for real-time control design and analysis.
A well known, and more importantly, physiologically verified model originated from the
work of Bergman et al. [15]. It utilises the concept of a remote compartment for the storage of
insulin to account for the time delay between injection of insulin and its utilization to reduce
blood glucose levels. Equations (1)–(3) show the equations used to define the system.
6
( ) ( )
( ) ( ) )3(/
)2(
)1(
32
1
IB
B
VtuIInI
IpXpX
tPGGXGpG
++−=
+−=
++−−=
�
�
�
where :
G = concentration of the plasma glucose above the basal level. (mMol L-1)
GB = basal level for plasma glucose concentration (mMol L-1), i.e. G + GB is
the total glucose in the blood plasma, where GB = 4.5 typically
X = utilization effect of insulin in a remote compartment (min-1)
I = concentration of the plasma insulin above basal level (mU L-1)
IB = basal level for plasma insulin concentration (mU L-1)
P(t) = exogenous glucose infusion rate (mMol L-1 min-1)
u(t) = exogenous insulin infusion rate (mU L-1 min-1)
p3 = subject dependant model parameter (mU-1 L min-2)
VI = insulin distribution volume (L)
n, p1, p2 = subject dependant model parameters (min-1)
The parameters, p1, p2 and p3, may be changed to represent different conditions of the glucose
regulatory system. For diabetic subjects:
p1 = 0, p2 = 0.025, p3 = 0.000013 (4)
These and other parameters are obtained from medical research [10,13,15].
The model in Equations (1)-(3) is a revised version of works by Bergman et al. [15] and
Furler et al. [14]. It was developed for modelling insulin sensitivity, a measure of how
7
efficiently the body responds to an insulin input after taking a glucose tolerance test. The
model is simple, yet accurately represents the essential dynamics of the human glucose
regulatory system for a variety of patients.
The three equations represent insulin production and infusion, insulin storage in a remote
compartment, and glucose input and insulin utilization in a second compartment. Insulin
inputs are either endogenous from the pancreas, or exogenous from a pump or injection and
are represented in Equation (3). Equation (2) defines the dynamics and delay in the transport
of insulin from the subcutaneous layer to the blood plasma and subsequent utilization.
Equation (1) represents the glucose levels in the blood stream and the dynamics of its reaction
with insulin. Figure 1 graphically outlines the glucose regulatory dynamics for the model of
Equations (1)-(3).
The values of n, VI, GB, IB employed for all simulations are defined, for an average-weighted
male, as follows [10]:
VI = 12 L, n = 5/54 min-1, GB = 4.5 mmol L-1, IB = 15 mU L-1 (5)
The controller uses a simple feedback loop that employs the blood glucose level above basal,
G, and its’ derivate, , as sensor inputs, and the exogenous insulin infusion rate, u(t), as the
control output. Glucose deviation from basal levels, G, and its slope, , are the only data
readily available to control the system. As a result, the controller measures the output from
Equation (1) while directly influencing the dynamics in Equation (3) via the control action. In
between are the time delay and additional dynamics of the remote compartment represented in
Equation (2).
8
To verify the model the controller output is set to zero, with p1=0.028, p2=0.025 and
p3=0.000013 to model normal human response [8], and simulated with an initial plasma
glucose level of G(0) = 15 mmol/L including the basal contribution. The results shown in
Figure 2 match the results obtained for a normal subject [11, 14]. The insulin infusion plot
represents insulin produced by the pancreas as the controller output is zero and normal human
response is enabled by the parameters selected. Note that the Plasma Glucose Concentration
graph shows the amount of glucose in the blood in units of mmol/L of glucose above the basal
level. The basal level is the fasting amount of glucose typically maintained in the absence of
additional glucose input. As expected, Figure 2 shows that when a normal person has blood
glucose above the basal level the pancreas acts to produce insulin to return the blood sugar to
the basal level. The transition is smooth over approximately 180 minutes (3 hours) with no
hypoglycemic overshoot.
3. DERIVATION OF CONTROL SCHEMES
In this section, three control schemes are presented, one closed-form semi-optimal solution
and another two input-output forms where the control input is not a direct function of the
model parameters, but determined by measured outputs.
There are many complex influences between glucose and insulin concentration for any
person, normal or diabetic. However, the steady state glucose concentration in the body is
mainly determined by how much insulin is present. In order to lower glucose concentration in
the blood, insulin needs to be released, injected or infused. Hence, the controller defines the
insulin infusion rate, u(t), based on the measured inputs G and G� . The goal is to minimise
9
excess glucose, G, and its rate of change, G� , and ensure that excursions from the basal value,
Gb, are minimised in magnitude and duration with no hypoglycemic overshoot below the
basal level.
Sustained blood glucose levels of around 8 mmol/L, or more, can result in significant
physiological damage. A value below GB can lead to loss of consciousness and, if large
enough, possible brain damage. Therefore, blood sugar levels must be maintained in a tight
range around the basal level, minimizing the size and duration of excursions in either
direction. This approach matches the concept that the ideal blood glucose curve for a
reasonable glucose input is typically considered to be relatively, if not completely, flat [3].
3.1 Relative Proportional Control (RPC)
One of the better-known, more effective diabetes control systems is relative proportional
control (RPC) [10]. This basic, widely used form of proportional control law is based on the
idea of strictly limiting the absolute blood sugar level by applying resisting "forces" (insulin)
in weighted proportion to the magnitude of the excursion from the desired (basal) level. The
control inputs are defined:
( ) bb
nVIuG
Gutu =
+= 00 ,1 (6)
This controller is based on a relative proportional control (G/Gb) with a constant term. In this
case, there is no input based on the derivative, or rate of change, of blood glucose level G� .
10
Note that when G = Gb the blood sugar is at the desired level and the insulin infusion rate is
u0, the basal infusion rate necessary to maintain blood glucose at a constant level.
3.2 Optimal Control
Since the model employed represents the essential dynamics of the human glucose regulatory
system, the optimal solution for u(t) can be obtained analytically. First, Equation (3) is solved
and the exact solution for I(t) determined in terms of the exogenous insulin infusion u(t).
( ) tdtueV
eeItI
ttn
I
ntnt
B ′′+−= ∫ ′−
−
0
)(1)( (7)
Similarly, using Equation (2), the solution for X(t) is obtained in terms of I(t) and using
Equation (7) can be directly expressed in terms of u(t).
( ) ( ) ( ) { }∫ ′−′−
+−+−−
= ′−−′−−−−−t
ttpttn
I
tpBtpntB tdeetunpV
pe
p
Ipee
np
IptX
0
)()(
2
3
2
3
2
3 222 )(1)(
(8)
Since the goal is to minimize G and G� , the assumption of 0== GG � at steady state is
applied to Equation (1) to obtain a steady state optimal solution, leaving:
)()( tXGtP B= (9)
11
Substituting Equation (8) into Equation (9) and using the Laplace transform and its inverse, to
simplify the convolution integrals, the optimal steady state exogenous insulin infusion profile
is obtained:
[ ] 0223
)())0()()(()0()()( utPnpPtPpnPtPGp
Vtu
b
++++++= ���� (10)
The optimal solution includes first and second derivatives of the exogenous glucose input
P(t). However, it is unrealistic to implement since P(t) is not known a priori, and because it is
not practical to remove insulin (or add glucose) as is suggested if some terms become
negative. Equation (10) is also an explicit function of the time constants and other model
parameters subjecting it to potential modelling error. This solution does act as a benchmark
for the performance of other controllers.
Note that the steady-state insulin infusion without exogenous glucose input, u(t) = nVIIB = u0,
corresponds to the result obtained from Equation (10) if P(t) = 0. The optimal solution will
provide an optimal insulin infusion profile for any case where the initial conditions match the
steady state assumptions. However, the result also assumes continuous sensor measurement
and feedback, rather than discretely spaced readings, which will influence simulation results.
3.3 Heavy Derivative PD Controller
The most effective controller developed, of several approaches, is a basic PD controller where
heavy emphasis is placed on the derivative term.
12
( ) ( )GkGkutu dp�++= 10 (11)
This controller incorporates proportional and derivative control with independent weightings
and the basal infusion u0. Heavy emphasis implies that the derivative gain, kd, is significantly
larger than the proportional gain, kp. As a result, this controller focuses almost exclusively on
controlling the slope of the blood glucose curve rather than its absolute magnitude. This
approach is a far different one than normally taken and made possible by the emerging
capability to measure blood sugar far more regularly via semi- or non- invasive means.
There are several benefits provided by derivative emphasized control for this non-linear,
delayed system. Most importantly, a large increase in glucose is observed before a large
proportional term, enabling a faster response time for the controller. Derivative terms can be
susceptible to noise, but in this case with appropriate filtering and technology it should not be
an issue. More specifically, the use of low pass finite or infinite impulse response (FIR or IIR)
filters, along with a variety of other well known filtering methods can be used to reduce noise
arising from glucose measurement error. Such filtering should be relatively straightforward
given the well-characterized measurement errors for approved glucose measurement methods.
In addition, the sampling period, even at 1-20 minutes, is long enough to make additional
measurements and average results if necessary. The only potential significant issue with
filtering over several samples is the small lag, or delay, that can occur in the filtered signal.
4. NUMERICAL TESTS AND ALGORITHM VERIFICATION
This section presents results from several different simulation tests to determine the efficiency
and effectiveness of these controllers in limiting blood glucose excursion and duration.
13
4.1 Oral Glucose Tolerance Test
The oral glucose tolerance test (OGTT) is a large step input test of the glucose regulatory
system often performed to diagnose diabetes. A fasting subject consumes 400–800
kilocalories of glucose and the response is observed. This beta-cell function test represents a
significant challenge to the pancreas.
The OGTT input may be modelled by the lognormal distribution curve defined in Equation
(2) and shown in Figure 3.
2))(ln()( cbtamePtP −−= (12)
Where Pm is the peak value and a, b and c are constants, which determine the slopes and
curvature. This function is employed because it is smooth, continuously differentiable, and
has zero initial conditions, making it easily implemented and physiologically representative.
It is easily modified to represent faster or slower absorption rates of exogenous glucose.
Figure 4 shows the glucose response for a OGTT using the PD controller with sensor
measurements every 1 and 20 minutes, and for the RPC system measuring every 20 minutes.
The nearly flat optimal control result is shown as well. There is also a graph of no controller,
labelled “Accumulation of glucose input”, which shows glucose input profile. Figure 5 shows
the insulin infusion rate for the RPC and PD controller, where the step shape is due to the 20
minutes between sensor measurements and changes to the insulin infusion (control) input.
14
Note that the insulin infusion rate is smoother and more like an injection for the PD controller
in both cases and particularly when sensor measurements are made every minute.
In judging controller performance the critical factors are the magnitude of the excursion for a
given input and the time required to return to basal blood sugar levels. The PD controller
limits the excursion and returns to the basal level much faster, in each case, due to the higher
infusion rates generated by large initial slopes of the glucose curve.
As seen in Figure 4, the optimal controller has an almost flat response. The insulin infusion
rate for the optimal solution is not shown, however, it is effectively an injection of
approximately 3 units of insulin followed by a small glucose infusion over 25 mins. Though
the optimal solution cannot be practically implemented, it does show that the optimal solution
approaches current injection practice in a way that achieves near perfect glucose response
given the controllers knowledge of the glucose input P(t).
Figure 5 also shows that as the sampling rate increases, the insulin infusion profile tends
towards injections with the optimal solution resulting in an injection of 3 units at the
beginning of the simulation. These results match current methods of managing diabetes that
have evolved over time, where insulin is injected before meals.
Sensor bandwidth, or sampling period, is varied since the current GlucoWatch measures blood
sugar levels every 20 minutes. The one-minute bandwidth graph shows what can be achieved
when future technology allows measurements at this rate. Smaller sampling periods have less
impact as the sampling period becomes small relative to system time constants that are larger
than 1 minute.
15
The Kp and Kd gains for the PD controller are set at different values based only on the sensor
bandwidth employed. The derivative term is emphasized due to its importance in enabling a
faster response to rising blood glucose levels. As expected, with increasing sampling time the
derivative emphasized PD controller performance deteriorates more than the relative
proportional control as derivatives lose some usefulness as the sampling period grows.
4.2 Multi-Meal Tests
The input for these simulations is designed to give the system a more strenuous physiological
test such as two meals within 6 hours. The inputs vary in magnitude from 50-400 calories and
are given in two groups, at t = 0,10,30 minutes and for the second meal at t = 210 and 300
minutes. At the end of 6 hours, the total intake of glucose into the body is over 1000 calories
with 1000 calories input over the first 4 hours. Figure 6 shows the glucose input profile for
these meals and the results are shown in Figures 7-9.
Figure 7 shows that the PD controller works slightly better than a normal person, in terms of
peak reduction as well as settling time with a sensor bandwidth of 1 minute. This result
illustrates the potential of this controller to replace physiological dysfunction. Figure 7 also
includes a plot of the response for a normal subject under the same conditions.
Figure 8 compares results at different sensor bandwidths and Figure 9 shows the insulin
infusion profiles. These figures show similar trends to the OGTT results, where the more the
insulin infusion profile tends towards injections, the better the control of blood sugar. For the
optimal controller, inset in Figure 9 with a scaled glucose input profile, the insulin input is
16
essentially two injections closely following the onset of each meal. The optimal controller
glucose curve in Figure 8 is nearly flat and as the sensor sampling period approaches zero this
solution should not vary at all to glucose inputs.
It is worth noting that the RPC performance is relatively resistant to changes in sampling
period from 1 minute to 20 minutes. The PD controller result almost doubles in magnitude
and requires slightly more time to return to basal level. The PD controller with a sample
period of 20 minutes also has a slightly hypoglycemic response where the graph dips just
below the basal value.
The primary results from this testing are that the algorithm is robust to significant multiple
inputs with the same gains as designed for OGTT inputs and that as sampling period
decreases the insulin input profile for the PD controller approaches that of injections at the
beginning of a meal input. This latter result tends to match current accepted clinical practice
for the management of Type I diabetes in New Zealand and elsewhere. In particular, the PD
controller with derivative weighting provides a basal input, u0, in the absence of elevated
blood sugar and leads to injection-like insulin infusion profiles when glucose inputs are
applied. In clinical practice low, basal insulin infusion is required to prevent diabetic
ketoacidosis (DKA) in the absence of insulin in the system, while injections of rapid acting
insulin are used to handle meals and other post-prandial glucose spikes. Basal levels are
typically maintained, clinically, via long-acting insulin injections. As a result, the behaviour
of the derivative weighted PD controller mimics current clinical practice that has evolved over
several years with a basal infusion level augmented in injection like insulin infusion profiles
for hyperglycemic events due to exogenous glucose input. Hence, there exists a great
similarity between the insulin infusion profile for the PD and optimal controllers, and typical
17
clinical recommendations using current practice. The primary difference is that these
controllers are able to effectively take advantage of a greater amount of blood glucose data
using a model-based controller.
4.3 Sensor Lag
A GlucoWatch measurement lags approximately one sample period due to the method by
which it obtains the value, and this lag is simulated with a 20 minute sampling period for the
PD controller to determine its impact on control efficacy, with results shown in Figure 10. As
expected, control response lags further allowing greater glucose excursions in both magnitude
and duration. The RPC is also shown for reference along with the zero-lag PD controller.
4.4 Sensor Failure
There is a possibility of sensor failure where sweat or other interference may cause a
measurement to be skipped. Figure 11 shows results for one and three successive sensor
failures, with the PD controller and a 20 minute sampling period, occurring 30 minutes after
the multi-meal simulation begins. The failures are timed for worst-case impact versus the
glucose input profile. A single sensor failure results in an acceptable range in glucose control,
however three successive failures results in a larger and longer excursion and greater
hypoglycemic overshoot. In both the sensor lag and sensor failure case the results clearly
indicate the over-reaction resulting from the delay in noting the rise in blood glucose level.
These results delineate how reliable the sensors must be to ensure safety and performance.
4.5 Comparison of Control Methods
18
It is seen that the heavy derivative PD controller outperforms the RPC. Figure 12 shows the
glucose response for the RPC controller with a variety of different proportional gains and a
20-minute sampling period. The RPC definition in Equation (6) has a gain of 1.0 on the
relative proportional term. However, as the gain is increased for better peak performance
hypoglycemic overshoot grows because even as the glucose levels are decreasing toward
basal a proportional controller is still infusing insulin. The result is hypoglycemia due to
over-insulinising a system with significant delay. Hence, simply increasing proportional gain
is not a practicable solution.
Figure 12 also shows the curve associated with the PD controller, which is able to command
significant insulin infusion earlier than a proportional dominated controller providing better
peak response control and limiting hypoglycemic overshoot. Due to the large derivative gain
the PD controller cuts the insulin infusion to zero when G� is significantly negative it allows a
more gentle drift back to the basal level when the glucose level starts dropping. This
approach enables better control and, with greater sensor bandwidth, drives the control input to
match current, effective diabetes management with infusion profiles that are essentially pre-
meal injections with a basal infusion rate at other times.
Overall, the control methods presented are found to be both simple and effective. This
control approach is less affected by patient specific parameters and as a result parameter
variation is better tolerated. For example, the PD controller is simulated with parameters for a
non-diabetic individual and the resulting exogenous insulin infusion is effectively zero.
Different types of insulin however cause difficulty where the control gains are tuned
19
assuming insulin that acts in about 12 minutes, with faster or slower acting insulin employed
in simulation.
5. CONCLUSIONS
The emergence of glucose sensors capable of providing blood glucose readings at a very high
rate is investigated for their ability to automate insulin infusion for diabetic individuals. Such
sensors open the possibility of controlling the slope of the glucose response curve, rather than
focusing strictly on absolute blood glucose level. Two controllers are derived, consisting of an
optimal solution and a simple PD controller where the derivative term heavily outweighs the
proportional term. A proportional controller is also simulated to highlight the effectiveness of
controlling the slope of the glucose curve rather than the absolute blood glucose level.
Oral glucose tolerance test and multi-meal inputs are employed to test the response of each
controller for a variety of sensor sampling periods. The impact of different sampling periods,
sensor lag, and the relative impact of proportional and derivative control is delineated to
determine the limiting values necessary for practical implementation. An optimal control law
is derived and used as a benchmark although it is not feasible for real-time implementation.
The optimal solutions, simulated with 1-minute sampling period, are essentially flat,
representing the limiting case for tight, active control of blood glucose levels.
The primary result is that a simple PD controller, heavily emphasising the derivative term, can
effectively and robustly manage a variety of glucose input profiles slightly better than even a
normal person. More specifically, the PD controller is shown to control glucose excursions up
to 50% better than proportional-based control schemes and slightly better than the normal
20
human system as represented by the model employed. Finally, as the sampling period drops,
the insulin infusion profiles for this type of PD controller mimic current diabetes management
utilizing pre-meal injections and low infusions at other times, verifying the basic approach.
As a result, far tighter control of blood glucose excursions from basal levels can be used to
reduce the impact of long-term exposure to (slightly) elevated blood glucose levels. In
addition, such an automated approach enables far more consistency in blood glucose
management, particularly for juveniles or others who may not be responsible for managing
their condition.
Further work is required to refine several aspects, including detailed study into robustly
handling different types of insulin. In addition, further simulation of these controllers using
more comprehensive models is required to further verify the methods developed. Finally,
examining more complex controllers will further improve the controlled response and
approach the optimal solution.
6. REFERENCES
[1] Diabetes New Zealand Inc. (2001). “Type II Diabetes: Managing for Better Health
Outcomes” PriceWaterHouseCoopers Report. S. Thomson, Prof. D. Beaven, M.
Jamieson, S. Snively, A. Howl, A.Christophersen
[2] American Diabetes Association. “Economic consequences of Diabetes Mellitus in the
US in 1997”. Diabetes Care 1998 21:296-309
[3] Alberti KGMM & Zimmett P Z. Definition, “Diagnosis and classification of diabetes
Mellitus – provisional report of a WHO consultation group”. Diabetes Med. 1998
15:539-53
21
[4] Van den Berghe, G, Wouters, P. Weekers, F., Verwaest, C., Bruyninckx, F., Schietz, M.,
Vlasselaers, D., Ferdinance, P., Lauwers, P. and Bouillon, R. (2001). "Intensive insulin
therapy in critically ill patients," New England J. of Med., vol. 345(19), pp. 1359-1367.
[5] Garg, SK, Potts, RO, Ackerman, NR; Fermi, S, Tamada, JA and Chase, PH, (1999).
"Correlation of finger-stick blood glucose measurements with GlucoWatch® biographer
glucose results in young subjects with Type 1 diabetes," Diabetes Care 22:1708.
[6] Worthington, D. (1990). "Use of models in the self-management of insulin-dependent
diabetes mellitus," Computer Methods and Programs in Biomedicine, vol. 32(3-4), pp.
233-239.
[7] Parker, R., Doyle, F., Harting, J., and Peppas, N. (1996). "Model predictive control for
infusion pump insulin delivery," IEEE Conf. Of Engineering in Medicine and Biology,
Nov 3, pp. 1822-1823.
[8] Doyle, F., Dorski, C., Harting, J., and Peppas, N. (1995). "Control and modeling of drug
delivery devices for the treatment of diabetes," Proc. of the American Control
Conference, June 21-23, pp. 776-780,
[9] Ollerton, R. (1989). "Application of optimal control theory to diabetes mellitus," Intl.
Journal of Control, vol. 50(6), pp. 2503-2522.
[10] Kienitz, K. and Yoneyama, T. (1993). "Robust controller for insulin pumps based on H-
infinity theory," IEEE Trans. On Biomedical Engineering, vol. 40(11), pp. 1133-1137.
[11] Fisher, Michael E., (1991). “A Semiclosed-Loop Algorithm for the Control of Blood
Glucose Levels in Diabetics”, IEEE Transactions on Biomedical Engineering, vol 38,
No. 1 pp 57-61
22
[12] Trajonoski, Z. and Wach, P. (1995). "Evaluation of subcutaneous route for the closed-
loop control of insulin delivery: Numerical and experimental studies," IEEE Conf. on
Engineering in Medicine and Biology, Sep 20-23, pp. 1357-1358.
[13] Cobelli, C., Nucci, G., and Del Prato, S. (1999). "Physiological simulation model of the
glucose-insulin system," IEEE Conf. on Engineering in Medicine and Biology, Oct 16,
pp. 999.
[14] Furler, Stuart M., Kraegen, Edward W., Smallwood, Robert H., and Chisholm, Donald
J. (1985). “Blood Glucose Control by Intermittent Loop Closure in the Basal Mode:
computer Simulation Studies with a Diabetic Model”, Diabetes Care, Vol. 8, No. 6 pp
553 – 561.
[15] Bergman, Richard N., Finegood, D. T., and Ader, M., (1985) "Assessment of Insulin
Sensitivity in Vivo", Endocrine Reviews, vol 6, no 1, pp 45- 86.
[16] Kraegen, E. W. and Chisolm, D. J., (1984), "Insulin responses to carting profiles of
subcutaneous insulin infusion: Kinematics modelling studies", Diaetologia, vol. 26, pp
208-213.
[17] Cobelli, C., Federspil, G., Pacibi, G., Salvan, A., and Scandellari, S. (1982) "An
Integrated Mathematics Model of the Dynamics of Blood Glucose and Its Hormonal
Control" Mathematical Biosciences, vol 58 pp 27 - 60.
[18] Lehman, E. and Deutsch, T. (1996). "Computer assisted diabetes care: a 6-year
retrospective," Computer Methods and Programs in Biomedicine, vol 50, pp. 209-230.
[19] Schrezenmeir, J. (1991). "Computer aids in dietetics," Diabetes, Nutrition and
Metabolism, Supplement 1, pp. 175-191.
[20] Skyler, J., Skyler, D., Seigler, D. and Sullivan, M. (1981). "Algorithms for adjustment
of insulin dosage," Diabetes Care, vol 4, pp. 311-318.
23
[21] Lehman, E. and Deutsch, T. (1993). "AIDA: a mk. II automated insulin dosage advisor,"
Journal of Biomedical Engineering, vol 15, pp. 201-211.
[22] Boroujerdi, M., Williams, C., Carson, E., Piwernetz, K., Hepp, K., and Sonksen, P.
(1987). "A simulation approach for planning insulin regimes," Intl. Symposium on
Advanced Models for Therapy of Insulin Dependent Diabetes, pp. 41-46.
[23] Fletcher, L. and Soden, P. (1991). "Diet construction using linear programming,"
Diabetes, Nutrition and Metabolism, Supplement 1, pp. 169-174.
[24] Berger, M. and Rotbard D. (1991). "A pharmacodynamic approach to optimising insulin
therapy," Computer Methods and Programs in Biomedicine, vol 34, pp. 241-253.
[25] Parker, R., Doyle, F., Harting, J. and Peppas, N. (1996). "Model predictive control for
infusion pump insulin delivery," IEEE Conf on Eng in Med and Biology, pp. 1822-1823.
[26] Ollerton, R. (1989). "Application of optimal control theory to diabetes mellitus,"
International Journal of Control, vol 50(6), pp. 2503-2522.
[27] Naylor, J., Hodel, A., Morton, B., and Schumacher, D. (1995). "Automatic control
issues in the development of an artificial pancreas," Proc. ACC, pp. 771-775.
[28] Kienitz, K. and Yoneyama, T. (1993). "Robust controller for insulin pumps based on H-
infinity theory," IEEE Transactions on Biomedical Engineering, vol 40(11), pp. 1133-
1137.
24
Figure 1: Diagram of the modelled dynamics for the human glucose regulatory system
Pancreas (Produces endogenous insulin)
Exogenous Insulin Input (insulin injection, pump etc.)
Remote Compartment (This provides the time delay between insulin injection and the glucose utilization)
I
u(t)
Exogenous glucose input (food intake etc.)
Liver (Produces endogenous glucose)
Plasma (The utilization of the glucose by the insulin takes place here)
X
G P(t)
Insulin production plant represented by Equation (3)
Remote compartment insulin storage plant represented by Equation (2)
Glucose production and utilization represented by Equation (1)
25
Figure 2: Response for a non-diabetic individual with G(0)+Gb = 15 mmol/L.
26
0 50 100 150 200 250 300 3500
0.1
0.2
0.3
0.4
0.5
GTT Input Profile
Time (mins)
Glu
cose
Inf
usio
n R
ate
(mm
ol/L
/min
)
P(t) = P*exp(-1*((log(0.5*t)-2)2))
Figure 3: Glucose input model of OGTT curve used in simulations.
27
Figure 4: Glucose response for an OGTT.
28
Figure 5: Insulin infusion rate for an OGTT
29
Figure 6: Multi-meal glucose input profile.
30
0 100 200 300 400 500 600 -5
0
5
10
15
20
25
30
35
40
45
Time (mins)
Net Glucose Input
Relative Proportional Control - 1min
PD - 1min
Normal Person
Figure 7: Glucose level of a normal human, and controlled diabetic individual with 1 minute sampling period for each controller type.
31
0 50 100 150 200 250 300 350 400 450 500 -5
0
5
10
15
20
25
30
35
40
45
Time (mins)
RPC - 1min sensor RPC - 20min sensor PD - 1min sensor
PD - 20min sensor
Optimal Steady State Infusion
Figure 8: Multi-meal glucose level for controllers at different sampling periods.
32
0 50 100 150 200 250 300 350 400 450 500 0
5
10
15
20
25
30
35
40
Time (mins)
0 50 100 150 200 250 300 -100
0
100
200
300
400
500 Optimal Case
PD - 20min sensor PD - 1min sensor
RPC - 1min RPC - 20min
Scaled Glucose Inputs at t=0,10,30,210,300 of P=0.2,0.1,0.4,0.5,0.05
Figure 9: Infusion rate for controllers at different sampling periods for the multi-meal test.
33
0 50 100 150 200 250 300 350 400 -5
0
5
10
15
20
Time (min)
20 min sensor lag PD
No sensor lag PD
20 min sensor lag RPC
Figure 10: Glucose response with a 20 min sensor lag
34
0 50 100 150 200 250 300 350 400-5
0
5
10
15
20Sensor Failure at 30 min for 20 and 60 mins
Time (min)
No failure
Failure for 20 mins
Failure for 60 mins
Figure 11: Results for the PD controller with 1 and 3 successive sensor failures, at 30 minutes, in the multi-meal test case.
35
Figure 12: Glucose response of multi-meal testing for derivative and proportional controllers.
Kp = 12
Kp = 6
Kp = 1
Kp = 0.5
Kp = 0.8
Kd= 0.8